This means that usage of a `StagingBelt` only requires passing along
the current `CommandEncoder` instead of both the encoder and the device.
It also makes the API reflect the fact that a `StagingBelt` cannot be
reused with a fresh device.
As far as I know, the only reason this was not done previously was that
`Device` was not clonable, and so doing this would have required an
`Arc<Device>` or other specific mechanism.